Network Layer Support for Gigabit TCP Flows in Wireless Mesh Networks This work is collaborated with Prof. Ramanathan, Univ. of Wisconsin Madison Chin-ya Huang Feb. 6, 2015 1 BBBBBBBBBBBBBBBB
Outline • Motivation • Challenges – Dynamic change of network environment – Related works • Spare Bandwidth Rate Adaptive Network Coding (SRNC) Scheme – Digraph diversity – Inter-flow network coding – Spare bandwidth exploitation – Buffer management • Evaluation • Conclusions and Discussions – Summary of SRNC – Extended research of SRNC 2 2
Motivation • Growing need of real-time streaming applications : – HDTV – Video conferencing – Movie on demands • Characteristics of these applications [Goel et al, TOMCCAP 2008] : – Long-lived – High data rate for each application (>1Gbps in the future) • Seamless deliver these applications require – High physical layer data rate (>1Gbps) – New techniques and protocols at higher layers 3 3
Problems of Mesh Networks • Mesh networks are able to provide high data link rate – Gigamesh network provides gigabit physical layer data rate in mesh networks – When data rate exceeds Gbps, packet loss, handoff delays, re-routing, etc. would have more severe impact on network throughput 4 4
Impact of different link bandwidth • TCP throughput drops while the link bandwidth fluctuates. – When the link bandwidth exceeds 1Gpb, TCP throughput < 50% of the maxflow. NewReno CTCP CBR 1.2 Normalized TCP Throughput 0.99 0.99 0.99 0.99 0.99 C B 1 0.75 0.8 0.74 0.63 10 TCP G 10 TCP 0.59 A 0.6 flows flows 0.46 0.42 0.4 0.31 0.27 0.21 0.21 I H 0.2 Cross traffic 0 10M 100M 500M 800M 1G Bottleneck link bandwidth (bps) 5 5
Spare Bandwidth Rate Adaptive Network Coding (SRNC) [Huang et al, TMC 2014] Goal: To develop strategies for improving end-to-end throughput – TCP based Develop and integrate the following four ideas: – Digraph diversity (direct acyclic graph diversity) – Inter-flow network coding – Spare bandwidth exploitation – Buffer management 6 6
Related Work: multipath routing • Forward packets through multiple routes [Wang et al, CoRoNet 2009] F B Pro: + More tolerant congestion + Load balance A H E C Con: - Routing is more G expensive - Fully exploit rerouting 7 7
Related Work: forward error correction (FEC) • Exploit available bandwidth by sending redundant packets to 2 recover packet loss [Medard F et al, INFOCOM 2009] B 1 1 1 1 2 2 + Loss tolerance 3 3 E A H - Require feedback of network condition 2 2 - Require sophisticated 1 2 1 C 2 schemes to adaptively G exploit available bandwidth - Increase packet 3 losses due to buffer overflow 8 8
Motivation of SRNC • SRNC: Adapt the idea of network coding into unicast transmission along with previous ideas • Network coding: linear combination of packets in the queue • Network coding is first proposed for multicasting aimed to increase datarate in 2000 [Ahlswede et al, TIT 2000] . • Packets are encoded from the same flow [Medard et al, INFOCOM 2009] . • Specifically, we network encode packets from multiple flows. We will show that: + Adapts to the changes in available bandwidth in a distributed fashion + Needs much less re-routing 9 9
Outline • Motivation • Challenges – Dynamic change of network environment – Related works • Spare Bandwidth Rate Adaptive Network Coding (SRNC) Scheme – Digraph diversity – Inter-flow network coding – Spare bandwidth exploitation – Buffer management • Evaluation • Conclusions and Discussions – Summary of SRNC – Extended research of SRNC 10 10
SRNC: digraph diversity • Gateway node (A) finds the direct graph in the network for multipath routing based on the destination (D1, D2). 2 E S1 B 1 2 1 D1 3 G A D 2 1 D2 S2 2 C F 3 11 11
SRNC: inter-flow network coding • Packets are encoded from multiple flows. • All mesh nodes network encode packets before forwarding them. - Loss of one packet may result in loss of all coded packets. 2 E S1 B 1 2 2 D1 G 3 A D D2 2 2 2 C F S2 3 12 12
SRNC: exploit available bandwidth • Assume all traffic from A to G without cross traffic • Each mesh node conditionally sends as many as possible to next hop – Send additional packets to outgoing links in a distributed fashion 2 E S1 B 1 2 2 D1 G 3 A D D2 2 2 2 C F S2 3 13 13
SRNC: buffer management • Gateway marks appropriate number of packets as high priority • Each mesh node adaptively manages packets in the queue. – High priority packets are sent first – Forward high priority packets no more than received – Drop low priority from buffer if necessary L L H H H L L H H H b) Buffer overflow occurs a) More bandwidth on outgoing links 14 14
SRNC: example • Although packet loss occurs, packets are still decodable because SRNC fully and distribultedly utilize the available bandwidth for transmission aimed to recover the packet loss during transmission. Packets are received successfully! 2 E S1 B 1 2 2 D1 G 3 A D D2 2 2 2 C F S2 3 15 15
SRNC • Packets are routed in a routing digraph. • Packets across different flows are linearly combined before being forwarded to each mesh node’s outgoing links. • Each mesh node forwards network encoded packets based on the available bandwidth. – Number of forwarded packets is not necessary to be the same as received. – Packets are discarded conditionally when necessary. • Access node reconstructs packets after decoding. 2 B S1 C 1 2 2 D1 A 3 G H D2 2 2 2 F 16 E S2 3 16
Outline • Motivation • Challenges – Dynamic change of network environment – Related works • Spare Bandwidth Rate Adaptive Network Coding (SRNC) Scheme – Inter-flow network coding – Diagraph routing – Forward error correction – Buffer management • Evaluation • Conclusions and Discussions – Summary of SRNC – Extended research of SRNC 17 17
Simulation Configurations Simulation tool: Network Simlator-2 (NS-2) Simulation topology: x (Mbps) : Average link bandwidth in the mesh. y (Mbps) : Link bandwidth to the users. x C x D x B x x y Internet G x y A y H J x x y I x : Mesh node x : TCP sender : Wireless host 18 18
Impact of different link bandwidth • Maximum possible TCP Throughput = min {y, } • x = 10 Gbps MR PRO [JK2009] SRNC Average TCP Throughput (Mbps) 1000 Large Spare Moderate Spare Tight Spare 900 Bandwidth Bandwidth Bandwidth 800 700 600 500 400 300 200 100 0 500M 800M 1G 1.2G 1.5G Link bandwidth between mobile hosts and 19 19
Fairness between heterogeneous flow sets • TCP Cross-traffic does not implement SRNC. • The performance of cross-traffic is not effected by SRNC. • The performance of TCP in SRNC is improved after bandwidth fluctuation. – SRNC utilizes spare bandwidth to enhance the throughput. Before After 2 Ratio of SRNC over MR 1.5 Throughput 1 0.5 0 TCP Cross-traffic 20 20
Effectiveness in Different Topologies (b) Topo 2 (a) Topo 1 (d) Topo 4 (c) Topo 3 (e) Topo 5 (f) Topo 6 (g) Topo 7 : TCP flows : TCP Cross-traffic flows 21 21
Effectiveness in Different Topologies The achieved throughput for TCP and TCP cross-traffic before and after bandwidth fluctuation (Cross- (Cross- (TCP,Before) (TCP,After) traffic,Before) traffic,After) Topo 1 0.93 1.98 1 1.05 Topo 2 0.95 5.25 1 1.04 Topo 3 0.95 2.15 1 1 Topo 4 0.9 3.61 1 1.09 Topo 5 0.93 9.11 1 1.36 Topo 6 0.93 4.23 1 1.48 Topo 7 0.93 0.72 1 1.43 22 22
Impact of wireless loss • Consider wireless loss occurs on each link with independent probability p . • Approximately, the wireless loss of the mesh is 4p. PRO [JK2009] SRNC 700 TCP Throughput (Mbps) 600 500 400 300 200 100 0 0 2 4 6 8 10 Overall Wireless Packet Loss Probability (%) 23 23
Outline • Motivation • Challenges – Dynamic change of network environment – Related works • Spare Bandwidth Rate Adaptive Network Coding (SRNC) Scheme – Inter-flow network coding – Diagraph routing – Forward error correction – Buffer management • Evaluation • Conclusions and Discussions – Summary of SRNC – Extended research of SRNC 24 24
Summary • Link bandwidth fluctuates due to the change of network environments. – TCP throughput drops significantly when BDP increases. – Network needs to react to the change to sustain the performance. • SRNC increases TCP throughput by effectively utilizing available bandwidth in the network. – Four components are involved. • SRNC shares bandwidth fairly with cross traffic no matter SRNC is applied or not. • SRNC is fully distributed and can be integrated in existing protocol stack and deploying in real network testbed. 25 25
Experimenting with SRNC • To execute the network coding in a link rate – Implement on FPGA • Global Environment for Network Innovations (GENI) project : – Implementing SRNC in Openflow enabled router. – Evaluating SRNC in network testbed. Internet Internet 26 26
Recommend
More recommend
